The Radeon HD 2900 XT is an immensely complex piece of hardware, with no fewer than 700 million transistors squeezed onto the die. Although AMD was keen to talk about how the HD 2000 series of chips will be based on a new 65nm process, the first GPU to break cover is actually based on an 80nm manufacturing process. I’m in no doubt that eventually the HD 2900 XT will transition to a 65nm process, and when that happens we could see increased clock speeds and lower power draw. As things stand, the GPU has a core clock speed of 742MHz, and the fact that the lower specced 65nm HD 2600 XT card has a core frequency of 800MHz, lends weight to the idea that we can expect higher clocks when the HD 2900 XT drops to a 65nm process.
With the X1000 series of Radeon cards ATI introduced its Ring Bus Memory Controller, which potentially allowed unprecedented memory bandwidth, assuming that you had fast enough memory to take advantage of it. Now AMD has evolved the Ring Bus controller and empowered the HD 2900 XT card with an astonishing 512-bit memory interface – double the bandwidth seen on any other card to date.
Considering the super-fast memory interface, it’s somewhat perplexing that AMD has decided to equip the first batch of HD 2000 XT cards with only 512MB of memory, when nVidia is already putting 640MB on its 8800 GTS parts and 768MB on the top end 8800 GTX parts. Even more confusing is that AMD/ATI delivered the first 1GB graphics card with the workstation incarnation of the X1800 XT - the FireGL V7350.
When questioned about the modest memory complement, AMD told me that board partners can specify as much memory as they like on their boards. Unfortunately that isn’t the case, since the first batch of cards will be reference boards manufactured by AMD, so the board partners will have to take what they’re given. That said, I did see sample cards out in Tunisia with 16 memory chips on them, which would obviously indicate 1GB rather than 512MB. AMD was however honest enough to tell me that the decision to go with only 512MB at launch was made in order to hit a certain, and undoubtedly lower price point.
The initial samples of the HD 2900 XT are equipped with GDDR3 memory, although you can expect a transition to GDDR4 at some point. The chips on the reference board are running at 825MHz (1650MHz effective) – although that’s quite high for GDDR3, the upper limit of GDDR3 frequency seems to be rising all the time, so there’s likely to be some headroom in there for overclockers.
The heart of the HD 2900 XT chip is an array of 320 stream processing units, which can turn their hand to pretty much anything. AMD has concentrated on efficient execution with the HD 2900 XT, rather than employing a bulldozer approach. The key is ensuring that each and every one of the superscalar stream processors is being used to maximum efficiency, in order to render each frame without wasting any resources. One of the keys to this efficient operation is a switch to a unified shader model...